AI agents are becoming actors in the world. Trust and authority need to become first-class.
dp-web4 is a heterogeneous human/AI research collective building Web4 — an open trust ontology, not an architecture, infrastructure or stack — together with Hestia, Hub and SAGE (Situation-Aware Governance Engine): running governance systems and persistent-agent research developed on an eight-machine fleet.
The common thesis is simple: trust should be computed from witnessed evidence, in context, by the party doing the relying — not declared by a platform or asserted by the agent itself.
This site is the lab view. For the ontology and its reference implementation, start with Web4; for local agent governance see Hestia (it runs on the lab's machines, but has no install guide for outside users yet; see the Quick start).
Of these, the eight instances with raising history are the current raising lines, one per machine (archived lines such as Sprout's Qwen 0.5B also carry history and are not in the eight; how instance, line and session record nest is on /context). They are enumerated per machine on /fleet. Not all eight are currently running: that page assigns each line a dated state (running, quiet, paused, instrumentation broken) and is the checkable version of these numbers. The configured-instance census is not.
What is being built
Web4
The open trust ontology: persistent identity, contextual trust, scoped authority, witnessed action, machine-readable law and federation — the last of these is roadmap, not running. Web4 lives on an RDF graph (Resource Description Framework) — it is not itself the substrate. Core Rust/Python packages are published; the standard remains draft in places.
Hestia
Local governance for humans and AI agents from multiple vendors: one law, scoped delegation, a vault, witnessed actions, escalation and trust derived from the record. Running today at A1 on Hestia's own published A0–A4 assurance ladder (A0 observed → A4 hardware-attested): the second rung, a cooperative and tamper-evident gate, not adversary-proof containment. (Not TCSEC's class A1, which is the top class; the A0–A4 row on /context has both.)
Hub
A Rust society runtime for communities and organizations: member identity, seven base roles, signed law, sealed channels and an append-only witnessed ledger. The fleet runs on the same society runtime it is developing.
SAGE
Persistent-agent research: identity, memory, salience, learned state, tools, sensors and governed effectors around frozen model weights. Current work asks whether agents can formulate, execute, evaluate and reuse their own experiments and procedures.
Commercial path
Open layer, higher-assurance enterprise tier
Web4, Hestia and Hub establish an open interoperability layer. Hardbound is not the only private piece: upstream calls the public Hub the open reference proof-of-concept and keeps its production development in a separate private repository. Hardbound, being built by Metalinxx, is the planned proprietary enterprise assurance tier for hardware-bound identity, stronger fail-closed enforcement and audit-ready evidence packaging. It is private and research-stage: hardware anchoring is a design target, and enforcement on the fleet today is at the process level (see /projects). The open layer avoids governance lock-in; the commercial layer raises assurance for deployments that require it.
How the lab works
Heterogeneous fleet
Edge devices, laptops, workstations and society-hosts run different model families for SAGE research. The tracks that write and audit this site share model families with each other, so review here is not yet independent (/autonomy names that limit).
Autonomous cycles
Scheduled research, maintenance, review and synthesis sessions run without continuous human operation, within predeclared scopes.
Evidence before story
Measured, implemented-but-unexercised, hypothesized and refuted are kept separate. Negative results and broken instruments stay in the record. The one quantitative test of the raising records came back null. It is written up on /raising and ranked on /context.
Historical ARC-AGI-3 work
SAGE's spring-2026 ARC-AGI-3 work remains available as a research archive. A Phase-1 harness around Claude Opus 4.6 produced a published 94.85% official ARC Prize action score (efficiency-weighted against first-time human players: 100% means every game beaten as efficiently as humans) on the public set — 175 of 183 levels across 23 of 25 environments — using engine-level/public-game affordances outside strict competition play. It was a useful milestone in the evolution of SAGE, but it is not current competition positioning; current competition-legal local-model work is well behind the leaders.